Bangla speech recognition using 1D CNN and LSTM with different dimension reduction techniques

This paper presents a model of Bangla speech recognition using machine learning algorithms. Mel-frequency Cepstral Coefficient (MFCC) and Mel Spectrogram are extracted from a Bangla dataset. Commonly used dimension reduction techniques, Principal Component Analysis (PCA), Kernel-PCA (k-PCA), and T-distributed Stochastic Neighbor Embedding (t-SNE) are applied to the extracted features as a dimension reduction technique. At the end, as a classification tool, 1-dimensional Convolutional Neural Network (1D-CNN) and Long-Short Term Memory (LSTM) are utilized. Experimental results demonstrate that among the dimension reduction techniques, PCA demonstrates comparatively higher accuracy than the other state-of-art models by exhibiting 94.58% and 83.12% accuracy for 1D CNN and LSTM, respectively. In addition, it has been observed that dimension reduction techniques have no positive impact on 1D-CNN and LSTM. Without any dimension reduction technique, MFCC with 1D-CNN has demonstrated better accuracy compared to MFCC with LSTM by showing 97.26% and 93.83% of accuracy, respectively.